Orca Security is warning that enterprise AI adoption is moving faster than basic remediation. According to media coverage of a new company report, 99.9% of fixable AI vulnerabilities remain unpatched even as organizations push more AI systems into production. The headline figure, reported by Help Net Security and AiThority, points to a familiar problem in cloud security now spilling into AI stacks: teams can identify issues, but patching and operational follow-through are lagging.

That matters because AI systems are no longer confined to pilots or internal demos. As companies deploy models, data pipelines, vector stores, and AI agents into customer-facing and revenue-linked workflows, the security cost of delay rises. A vulnerability in an experimental lab environment is one thing; the same weakness in a production service tied to sensitive enterprise data is another.

What Orca Security is claiming

The central news event is the release of an Orca Security report, cited by AiThority as finding that 99.9% of fixable AI vulnerabilities remain unpatched as AI moves into production. Help Net Security separately highlighted the same topline conclusion. The available source evidence in this story cluster is limited to those reports and summaries rather than the full underlying research document, so some important details are not visible, including methodology, sample size, time period, and Orca Security’s precise definition of an AI vulnerability.

Even with those caveats, the reported finding is significant because it shifts the discussion from theoretical AI risk to operational hygiene. The issue is not only whether organizations have exposures in their AI environments, but whether they are actually fixing known, remediable issues after discovery.

That distinction is critical for enterprise AI. Many AI security conversations still center on frontier-model dangers, jailbreaks, or abstract governance concerns. Orca Security’s reported finding suggests a more immediate and practical gap: routine patching discipline is not keeping pace with enterprise AI deployment.

The report’s framing also implies that AI infrastructure is increasingly being treated as part of the broader cloud estate. That puts AI services, dependencies, and supporting components in the same risk category as other production systems, where unpatched flaws can become an entry point for data access, lateral movement, or service disruption.

Why unpatched AI vulnerabilities matter now

The timing of this warning reflects where the market is. Enterprises are moving from experimentation to integration, embedding AI into internal search, coding workflows, customer support, security operations, and productivity tools. That creates a larger attack surface spanning models, APIs, open source packages, orchestration frameworks, storage layers, and runtime environments.

When teams rush to ship AI features, security processes often arrive later. Builders may prioritize model quality, latency, and cost over patch management for supporting components. Product teams may assume that using a managed model endpoint reduces their exposure, even though the surrounding system still includes application code, connectors, permissions, and data paths that can introduce risk.

This is especially relevant for organizations building with AI agents. Agentic systems typically interact with multiple tools and services, which can multiply the number of dependencies and credentials in play. If patching practices are already weak in standard cloud environments, the challenge can become harder as AI agents gain broader access to enterprise systems.

For enterprise AI buyers, the report’s message is less about panic than about maturity. AI security is not only a model evaluation problem. It is also a software maintenance problem, a cloud visibility problem, and a governance problem. If vulnerabilities are fixable but still left open, the bottleneck may be process rather than detection.

A broader signal about AI in production

The most important part of the Orca Security report may be its subtitle, as relayed by AiThority: AI is moving into production. That phrase captures why this finding deserves attention from founders and platform teams.

In the pilot phase, organizations can tolerate rough edges. In production, they cannot. Once AI is tied to regulated data, customer workflows, or developer infrastructure, patching delays become harder to justify. Security leaders then have to answer familiar questions in a new context: what assets are exposed, who owns remediation, how quickly can fixes be applied, and what business systems are affected if something goes wrong?

The report also lands during a period when many companies are trying to standardize AI tooling. Some are consolidating around a small set of approved model vendors while building internal controls for prompts, data access, and audit trails. Others are still in a fragmented phase, with different teams using separate frameworks and services. In the latter case, patching can be particularly messy because no single owner has a full inventory.

That makes visibility central. If AI workloads are spread across cloud accounts, open source components, notebooks, APIs, and managed services, organizations may struggle to know where vulnerabilities exist, let alone remediate them quickly. In practice, the challenge is not just identifying flaws in AI systems but mapping AI systems into existing security and DevSecOps processes.

Evidence, limitations, and what is still unclear

The evidence available for this article comes from two media reports summarizing an Orca Security report. AiThority attributes the 99.9% figure directly to Orca Security, and Help Net Security independently surfaced the same claim. However, neither source excerpt included the underlying dataset, benchmark design, or full report text.

That means the strongest claims here should be treated as vendor-reported. Orca Security is a security company, and its research likely draws from environments it can observe through its own platform or customer base. Without the full report, it is not possible to verify how broadly the findings generalize across the market.

Several questions remain unanswered in the provided evidence. It is unclear what types of vulnerabilities were counted, whether the vulnerabilities were found in models themselves or in surrounding infrastructure, how Orca Security defined “fixable,” and over what time window patch status was measured. It is also unclear whether the figure refers to individual CVEs, cloud misconfigurations, package issues, or a combined set of exposures.

Those gaps do not invalidate the report, but they do affect how the number should be interpreted. A dramatic percentage can describe a serious market problem, or it can reflect a narrow dataset with specific assumptions. Buyers evaluating Orca Security or comparing findings across security vendors should look for methodological detail before treating the 99.9% number as a universal benchmark.

Still, even with limited transparency in the available sources, the direction of the finding aligns with what many security teams already report anecdotally: enterprise AI is being adopted quickly, while remediation workflows remain uneven.

What this means for builders and enterprise teams

For builders, the immediate implication is that AI features need the same operational discipline as any other production service. Teams using OpenAI, Anthropic, Microsoft Azure, or AWS to power applications may think primarily about model selection and inference cost, but patching responsibilities do not disappear when a third-party model is involved. The application layer, data connectors, retrieval systems, deployment images, and CI/CD pipelines still need standard security maintenance.

For platform owners, the report is a reminder to bring AI workloads under existing cloud security policies rather than managing them as exceptions. That includes asset inventory, vulnerability scanning, ownership tagging, remediation SLAs, and secrets management. If a company has separate controls for cloud and AI, the seams between them may become the weak point.

For enterprise buyers, procurement questions should move beyond model performance. When assessing vendors such as Orca Security or adjacent AI security platforms, buyers should ask how AI assets are discovered, how exposures are prioritized, and how remediation is tracked. Security posture matters more when AI systems touch sensitive data or execute actions through AI agents.

For founders, there is also a product lesson. Shipping AI quickly can help win market share, but enterprise customers will increasingly ask how security is maintained once systems reach production. Startups selling into regulated industries may need to show not just red-team results and policy controls, but evidence of repeatable patching and software supply chain hygiene.

What to watch next

The first thing to watch is whether Orca Security publishes fuller methodology or additional detail around the report. Sample size, environment mix, and category definitions will determine how seriously security leaders can benchmark themselves against the 99.9% figure.

Second, watch whether other security vendors report similar patching gaps in enterprise AI environments. If multiple firms converge on comparable results, the market will have a stronger case that this is a systemic issue rather than a vendor-specific dataset.

Third, track how enterprise AI governance programs evolve. If companies start folding model services, retrieval stacks, and AI agents into mainstream DevSecOps dashboards, patching rates could improve. If AI remains operationally separate, remediation gaps may persist.

Finally, expect buyers to scrutinize production-readiness claims more closely. As enterprise AI deployments mature, questions about visibility, patch velocity, and accountability may become as important as benchmark performance or demo quality.

Creati.ai perspective

The sharpest takeaway from the Orca Security report is not the exact percentage, which still needs fuller methodological context, but the category of failure it highlights. The weak point appears to be mundane execution: known issues in production AI environments are not being fixed fast enough. That is a more immediate commercial and operational problem than many of the headline-grabbing debates around advanced model behavior.

For the AI market, this is a sign that enterprise AI is entering a more practical phase. Buyers no longer need only better models; they need systems that fit into real security operations. Companies that can connect AI development to existing cloud controls, prove remediation discipline, and reduce the sprawl around AI agents will be in a stronger position as enterprise AI spending shifts from experimentation to durable platforms.